Data-driven Model Reduction for Soft Robots via Lagrangian Operator Inference

Harsh Sharma, Iman Adibnazari, Jacobo Cervera-Torralba, Michael T. Tolley, Boris Krämer

Abstract

Data-driven model reduction methods provide a nonintrusive way of constructing computationally efficient surrogates of high-fidelity models for real-time control of soft robots. This work leverages the Lagrangian nature of the model equations to derive structure-preserving linear reduced-order models via Lagrangian Operator Inference and compares their performance with prominent linear model reduction techniques through an anguilliform swimming soft robot model example with 231,336 degrees of freedom. The case studies demonstrate that preserving the underlying Lagrangian structure leads to learned models with higher predictive accuracy and robustness to unseen inputs.

Introduction: The Challenge of Soft Robot Modeling

Soft robots are notoriously difficult to model because they are inherently infinite-dimensional and exhibit complex nonlinear dynamics. Currently, engineers face a significant trade-off between simplified physics models, which are fast but often inaccurate, and high-fidelity Finite Element Method (FEM) simulations. While FEM models are highly accurate, they are computationally "heavy"—the model used in this study features over 231,000 degrees of freedom (DOF), making it impractical for real-time control.

Furthermore, many existing Model Order Reduction (MOR) techniques are intrusive, requiring access to the internal source code and operators of the simulator. This research seeks a nonintrusive method to build fast, accurate models using only simulation data while ensuring the resulting model adheres to physical principles.

Methodology: Lagrangian Operator Inference (LOpInf)

The core technique proposed is Lagrangian Operator Inference (LOpInf). This approach combines data-driven flexibility with the rigorous constraints of classical physics. The process involves:

  1. Data Collection: Running a high-fidelity FEM simulation (the Full-Order Model or FOM) to collect "snapshots" of displacements, velocities, and accelerations.
  2. Dimensionality Reduction: Utilizing Singular Value Decomposition (SVD) to identify a low-dimensional subspace capturing the majority of the system's energy.
  3. Structure-Preserving Learning: Solving a constrained optimization problem that forces learned matrices (Mass, Damping, Stiffness) to be symmetric and positive-definite, ensuring the model obeys Euler-Lagrange equations.

The authors compare LOpInf against two popular "black-box" methods: Dynamic Mode Decomposition with control (DMDc) and the Eigensystem Realization Algorithm (ERA/OKID).

Experimental Setup: Anguilliform Soft Robot

The study utilizes a simulated eel-like swimming soft robot built within the SOFA (Simulation Open Framework Architecture) framework.

Schematic and 3D model of the anguilliform swimming soft robot used for data collection.

The robot model consists of 231,336 DOF and is actuated by six fluid elastomer actuators (FEAs). Training and testing data were generated from six simulation episodes using random control sequences, each consisting of 2,000 snapshots over a 2-second interval.

Key Results and Performance Comparison

The research demonstrates that LOpInf can achieve massive compression, reducing the system from over 230,000 DOF to just 10–16 DOF while maintaining high accuracy.

Comparison of relative output error for LOpInf (red), DMDc (green), and ERA/OKID (blue) across different reduced dimensions.

Key findings include:

  • Superior Extrapolation: In time extrapolation tests, LOpInf remained stable and accurate, whereas DMDc failed to generalize beyond the training window.
  • Robustness: When tested on completely new control sequences, LOpInf consistently achieved lower relative output errors than both DMDc and ERA/OKID.
  • Physical Consistency: By enforcing a Lagrangian structure, LOpInf provides a physical "guardrail" that prevents the overfitting common in purely data-driven models.

Conclusion and Future Work

While linear LOpInf models are highly effective for a wide range of motions, they eventually reach a "nonlinear wall" when faced with extreme geometric nonlinearities or complex twisting behaviors. Future work will focus on extending this framework to nonlinear model reduction, potentially incorporating Koopman operator theory or structure-preserving machine learning to handle high-strain regimes.

Summary updated

BibTeX
@inproceedings{sharma2024data,
  author = {Harsh Sharma and Iman Adibnazari and Jacobo Cervera-Torralba and Michael T. Tolley and Boris Krämer},
  title = {{Data-driven Model Reduction for Soft Robots via Lagrangian Operator Inference}},
  booktitle = {IFAC-PapersOnLine},
  year = {2024},
  doi = {10.1016/j.ifacol.2024.10.119},
}
Figure